companding_mu_law — ONED signal op

Data kinds: signalsignal

Call: import fullseye as fs; fs.ledger.companding_mu_law(x, mu=255.0, bits=8) (to call the implementation directly, import dsp; dsp.companding_mu_law(x, mu=255.0, bits=8); from the registry, ops1d.get("companding_mu_law"))

Usage

mu-law companding — the G.711 curve, used here on any 1-D signal.

Compress with `F(v) = sign(v) * ln(1 + mu|v|) / ln(1 + mu)` on the signal

scaled to `[-1, 1]`, quantise uniformly, expand back. The steps end up fine

near zero and coarse near full scale, which is the right allocation when the

interesting part of the signal is small compared with its peaks — speech,

vibration, anything with a large crest factor.

This is where mu-law comes from: the image operator

`companding_mu_law` is the same curve applied to intensity. Reporting both

keeps the family honest about which dimension the technique was designed for.

Applicability — it is the crest factor that decides. Measured on a sine

of amplitude *A* with one sample pinned at full scale, so the crest factor is

exactly `1/A` (mean square error relative to a uniform quantiser):

crest 50 4 bit 0.011 6 bit 0.014 (about 90x better)

crest 20 4 bit 0.122 6 bit 0.101

crest 6.7 4 bit 0.613 6 bit 0.616

crest 2.0 4 bit 4.22 6 bit 6.03 (several times WORSE)

So the rule is crest factor above roughly 7 — speech, vibration, impact.

Below that, a plain uniform quantiser wins and mu-law actively hurts.

★Note the peak is a single sample: on random signals of the same family

the advantage swung between 0.55 and 0.87 purely with the seed, because the

largest excursion sets the scale. Measure the crest factor of *your* signal,

do not assume it from the distribution.

Other limits: `mu` near 0 degenerates to uniform quantisation (that is how

you check the curve is doing anything), and the curve is fixed — unlike a

Lloyd-Max codebook fitted to the signal — which is the point when values must

stay comparable across recordings.

References (sample data, literature)

• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).

• Operator provenance and references — the sources of the research/methods this op family came from.

• The canonical algorithm (author, year) and its uses are named in the family usage guide above.

Runnable examples (verified samples that actually call this op)

gallery2d_gray_arithpy -3.11 examples/gallery2d_gray_arith.py

Ops the type connects to (they accept signal as input)

create_funct_1d_array · create_funct_1d_pairs · smooth_funct_1d_gauss · smooth_funct_1d_mean · derivate_funct_1d · integrate_funct_1d · zero_crossings_funct_1d · local_min_max_funct_1d

Same category (signal)

lowpass · highpass · bandpass · envelope · rms · local_std · quantize · resample


*Provenance: dsp.py — ONED operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*

© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.